Why unlearnable examples work: a new perspective from mutual information reduction

An arXiv paper explains why unlearnable examples work from the perspective of mutual information reduction and proposes MI-UE. The study finds that effective unlearnable examples reduce the mutual information between clean features and poisoned features, and that the deeper the network, the greater the improvement in unlearnability is accompanied by a decrease in mutual information; the authors further prove that minimizing the conditional covariance of intra-class poisoned features can reduce the mutual information between distributions. MI-UE reduces covariance by maximizing intra-class feature cosine similarity, and experiments show it significantly outperforms prior methods and remains effective under defense mechanisms.